Robot interaction with objects based on semantic information associated with embedding spaces

Abstract

Techniques described herein relate to using reduced-dimensionality embeddings generated from robot sensor data to identify predetermined semantic labels that guide robot interaction with objects. In various implementations, obtaining, from one or more sensors of a robot, sensor data that includes data indicative of an object observed in an environment in which the robot operates. The sensor data may be processed utilizing a first trained machine learning model to generate a first embedded feature vector that maps the data indicative of the object to an embedding space. Nearest neighbor(s) of the first embedded feature vector may be identified in the embedding space. Semantic label(s) may be identified based on the nearest neighbor(s). A given grasp option may be selected from enumerated grasp options previously associated with the semantic label(s). The robot may be operated to interact with the object based on the pose and using the given grasp option.